Feature association for object tracking

This paper addresses the problem of feature-based estimation of the 3D motion (rotational + translational) and structure of a rigid object from a sequence of 2D monocular images. A rigid object is represented by a set of junctions - groupings of line segments that meet at a single point - which has several advantages over other techniques. An overallscheme for 3D object tracking using enhanced junction detection via a modified Hough transform and state estimation using an unscented Kalman filter was developed in a previous paper. This paper focuses on a crucial part of the overall tracking scheme the association (matching) of the predicted and observed junctions. The problem is formulated systematically as a general global assignment problem which allows for the treatment of uncertainties such as occlusion and appearance of new features. Appropriate assignment costs are proposed that account for junction topology and geometry. In addition, a general convex programming approach for two and multiple frame junction matching is also proposed. Simulation results illustrating the accuracy of the junction association algorithms as well as the overall object tracking scheme are provided. Overall, the approach demonstrates that well established data association methods developed for "point" multitarget tracking can, after appropriate adaptation, be very useful for tracking rigid objects.